OpenMP并行块外的种族条件（ThreadSanitizer）；误报？

Question

下面的最小示例计算从1到1000的所有数字的总和，并与OpenMP并行化。

#include <iostream>

double sum;

void do_it() {
  const size_t n = 1000;
#pragma omp parallel
  {
#pragma omp for
    for (size_t i = 1; i <= n; ++i) {
#pragma omp atomic
      sum += static_cast<double>(i);
    }
  }
}

int main() {
  sum = 0.;
  do_it();
  std::cout << sum << std::endl;
  return 0;
}

我尝试使用clang++-6.0.0和g++-5.4.0以及ThreadSanitizer进行编译。两种编译器都会在libomp.so / libgomp.so中产生一些关于竞争条件的警告，我认为它们是误报，以下是关于我的代码的警告：

==================
WARNING: ThreadSanitizer: data race (pid=22081)
  Read of size 8 at 0x000001555f48 by main thread:
    #0 main /home/arekfu/src/foo/openmp.cc:20 (openmp+0x4be0ce)

  Previous atomic write of size 8 at 0x000001555f48 by thread T11:
    #0 __tsan_atomic64_compare_exchange_val ??:? (openmp+0x476470)
    #1 .omp_outlined._debug__ /home/arekfu/src/foo/openmp.cc:12 (openmp+0x4be011)
    #2 .omp_outlined. /home/arekfu/src/foo/openmp.cc:8 (openmp+0x4be011)
    #3 __kmp_invoke_microtask ??:? (libomp.so.5+0x994b2)

  Location is global '<null>' at 0x000000000000 (openmp+0x000001555f48)

  Thread T11 (tid=22093, running) created by main thread at:
    #0 pthread_create ??:? (openmp+0x4284db)
    #1 __kmpc_threadprivate_register_vec ??:? (libomp.so.5+0x5bc1f)
    #2 __libc_start_main /build/glibc-LK5gWL/glibc-2.23/csu/../csu/libc-start.c:291 (libc.so.6+0x2082f)

SUMMARY: ThreadSanitizer: data race /home/arekfu/src/foo/openmp.cc:20 in main
==================

虽然我在代码中看不到任何数据争用！

我也尝试用atomic部分替换critical更新，如下所示：

#pragma omp critical
      {
        sum += static_cast<double>(i);
      }

这将更改警告，但新警告没有多大意义：

==================
WARNING: ThreadSanitizer: data race (pid=27477)
  Write of size 8 at 0x000001555f48 by thread T4:
    #0 .omp_outlined._debug__ /home/arekfu/src/foo/openmp.cc:13 (openmp+0x4be0a2)
    #1 .omp_outlined. /home/arekfu/src/foo/openmp.cc:8 (openmp+0x4be0a2)
    #2 __kmp_invoke_microtask ??:? (libomp.so.5+0x994b2)

  Previous write of size 8 at 0x000001555f48 by thread T3:
    #0 .omp_outlined._debug__ /home/arekfu/src/foo/openmp.cc:13 (openmp+0x4be0a2)
    #1 .omp_outlined. /home/arekfu/src/foo/openmp.cc:8 (openmp+0x4be0a2)
    #2 __kmp_invoke_microtask ??:? (libomp.so.5+0x994b2)

  Location is global '<null>' at 0x000000000000 (openmp+0x000001555f48)

  Thread T4 (tid=27482, running) created by main thread at:
    #0 pthread_create ??:? (openmp+0x42857b)
    #1 __kmpc_threadprivate_register_vec ??:? (libomp.so.5+0x5bc1f)
    #2 __libc_start_main /build/glibc-LK5gWL/glibc-2.23/csu/../csu/libc-start.c:291 (libc.so.6+0x2082f)

  Thread T3 (tid=27481, running) created by main thread at:
    #0 pthread_create ??:? (openmp+0x42857b)
    #1 __kmpc_threadprivate_register_vec ??:? (libomp.so.5+0x5bc1f)
    #2 __libc_start_main /build/glibc-LK5gWL/glibc-2.23/csu/../csu/libc-start.c:291 (libc.so.6+0x2082f)

SUMMARY: ThreadSanitizer: data race /home/arekfu/src/foo/openmp.cc:13 in .omp_outlined._debug__
==================

这些警告是否表示真实的数据竞争，或者它们是误报？

Answer 1

“问题”是第20行对sum的读取操作：

std::cout << sum << std::endl; // here you are reading the value of sum

TSAN无法推断此读取与循环中的（原子）更新之间的关系之前发生线程间事件。但是当然存在这种关系，因为所有线程都在omp块的末尾同步。是的，这是一个误报。

这篇文章提供了更多信息，如何使用OpenMP避免这种误报：Can I use Thread Sanitizer for OpenMP programs?